Decompile-Bench is the first open-source dataset comprising two million binary-source function pairs condensed from 100 million collected pairs. The dataset, created by LLM4Binary, includes both C and C++ source code, whereas earlier models were limited to C only. It was last updated on September 30, 2025.
Use Cases
- Training decompilation models based on binary-source function pairs.
- Benchmarking the performance of decompilation tools.
- Analyzing the relationship between compiled binaries and their original source code.
- Expanding decompilation support to include C++ based on the dataset's language coverage.
Strengths
- Contains two million binary-source function pairs.
- Derived from a collection of 100 million function pairs.
- Supports both C and C++ source code, expanding language coverage.
Limitations
- Column-level documentation is absent; field semantics must be inferred after download.
- Row count is unknown, which may limit suitability assessment.
- Description metadata is limited; actual data quality requires manual inspection after download.
Provenance
- Source
- LLM4Binary
- Collection Method
- Condensed from 100 million binary-source function pairs compiled from permissively licensed GitHub projects.
- Time Range
- null
- Freshness
- Last updated 2025-09-30 11:12:17; freshness should be verified.
- Geography
- null